Modeling the neural mechanisms underlying rule-based behavior

Sophisticated goal-directed behavior requires the memory of recent events, the knowledge of the context in which they occur, the goals we intend to reach, and the actions we can perform to obtain them. These are all elements characterizing our mental state. The execution of complex behavior can then be viewed as a flexible rule-based navigation of mental states. Here I develop a theoretical framework which assumes that mental states are neurally represented as stable patterns of sustained activity. I show that one of the conditions necessary to unify the stability of the actively maintained activity patterns with the possibility of a flexible event-driven switch, is the existence of an extensive number of neurons displaying mixed selectivity to conjunctions of mental states and external events. Remarkably, this particular selectivity can be naturally obtained through Randomly Connected Neurons (RCNs). I characterize the capacity, scaling, and response properties of several classes of networks displaying mixed selectivity through RCNs. This analysis suggests that a distributed heterogeneous mode of activity is the neural hallmark of a robust and computationally efficient execution of complex rule-based behavior. This motivates a detailed analysis aimed at characterizing the degree of mixed selectivity in neurophysiological recordings in the orbitofrontal cortex (OFC) and amygdala of monkeys executing a context-dependent trace conditioning task. These areas display extensive mixed selectivity to the task-relevant stimuli and the affective value they bear. Starting from this result, I proceed to illustrate a theory of how this kind of mixed selectivity neurons can accommodate the creation of new mental states encoding the behavioral context. This theory postulates the presence of two hierarchically organized learning systems. The first one rapidly learns the value of the presented stimuli on a feedback-based manner. It then relays this information to a slower learning system which displays mixed selectivity to the stimuli and their value. This system modifies the synaptic connections between neurons on the basis of the temporal contiguity of their activation. The result of these synaptic modifications is the creation of new patterns of reverberating activity encoding the temporal information of the task contingencies. In turn, these new self-sustained activity patterns participate in sequences of activation on longer time-scales and mediate the iterative creation of new patterns of self-sustained activity. This process of fusion of self-sustained activity patterns is termed attractor concretion. The mental states obtained as a result of attractor concretion represent the temporal context on relevant behavioral time-scales and contain information allowing an animal to unambiguously assign a value to the events that initially appeared in different situations with different meanings, thereby contributing to goal-directed behavior.